Background: Identifying drug-target interactions (DTIs) is a central focus in pharmaceutical research, as accurately recognizing these interactions can play a crucial role in developing modern and targeted therapies. In recent years, numerous deep learning-based models have been introduced to predict these interactions. However, several challenges remain. Existing methods often fail to incorporate the three-dimensional structures of drugs and proteins alongside their SMILES and FASTA sequences, resulting in lower prediction accuracy. Furthermore, many approaches utilize only partial sequence data, thereby overlooking critical information. This lack of spatial and comprehensive sequence awareness ultimately limits the accurate modeling of molecular interactions and binding mechanisms.
Methods: In this study, we introduced TGATS2S-v1 and TGATS2S-v2, two novel deep learning frameworks designed to address the critical challenge of Drug-Target Interaction (DTI) prediction by integrating 3D structural information of both drugs and target proteins alongside their canonical sequence representations (SMILES and FASTA). The proposed methods leveraged three-dimensional structural information to enhance DTI prediction and were tested on the Davis dataset.
Results: The results of the proposed methods were thoroughly analyzed. By integrating 3D structural data, the predictive power of the models improved significantly. Evaluations showed that these models consistently outperformed advanced baseline models, delivering higher accuracy and robustness in all cases. The proposed model achieves state-of-the-art performance, improving PR-AUC by over 20%.
Conclusion: These findings indicate that incorporating 3D structural information plays a pivotal role in overcoming the limitations of previous models and paves the way for the discovery of more effective drugs and therapeutic advancements.
